Surprise! Less memory might speed up photoshop and other programs.
See the article at
http://firingsquad.com/hardware/building_gaming_opteron_2003_Part2/page14.asp
They tested several platforms including a Pentium on a board with the
875P (Canterwood) chipset. They compared 2 memory sticks of 512MB
(1GB) to 4 memory sticks of 512MB (2GB). The memory was DDR400.
Their tests showed the 1GB system was faster, sometimes by a
significant margin.
At least one reason for the decreased performance is that the 875P
increases memory latency if you use more than 2 sticks of memory.
In a forum on the same site the article authors acknowledge that some
of the slowdown with 4 memory sticks might be specific to the
motherboard in their system. They were not using an ASUS board.
This article is of great interest to me since I'm planning to build a
system for running photoshop based on the ASUS P4C800-E Deluxe
motherboard which uses the 875P chipset.
Any comments on the above article? Some googling failed to find much
discussion based on the article's findings.
Anyone else out there have any data on photoshop performance on a 875P
system with 2 sticks of ram compared to 4 sticks?
If there really is a significant photoshop performance hit when
running 4 sticks of ram, then the best memory configuration for
photoshop might be 2 sticks of 1024 ram (2GB). This would be a
matched pair running in dual channel mode.
Joey
One of the problems with any review articles, is you don't know
how competent the reviewers are at configuring the system.
I have yet to see an article that compares all aspects of memory
performance on 875/865. Here are the issues that affect performance:
1) Clock rate. The single most important factor in processor
performance and memory performance. Clock rate is more important
than a low CAS for example.
2) Command rate. When the number of loads goes up on a memory
bus, command rate has to be switched from 1T to 2T. Basically,
an address is placed on the bus for two cycles, but the strobe
for the address is only active for the second cycle, when command
rate is set to 2T. This improves setup time under heavy address
loading. Two double sided modules on a single bus is 32 chips,
and generally requires 2T timing. Both the Athlon64 and the
Pentium systems will face this issue. In a way, this is like
a CAS penalty.
3) CAS number. There is a slight improvement with CAS2 over CAS3.
4) Intel has PAT, which is a shortening of the memory path when
the FSB is synchronous with the transfer rate from the memory
subsystem. There are a whole bunch of conditions that seem to
influence whether PAT exists or not. There is no single register
that I can find on a 875/865 Northbridge to identify whether
PAT is in action or not (unless it is an undocumented feature).
It is possible the effects of Command Rate are considered a
part of PAT, because I cannot find a BIOS setting for Command
Rate on the P4C800-e. This article is just the tip of the
iceberg:
http://www.anandtech.com/mb/showdoc.aspx?i=1851&p=5
5) The interleave factor. More banks is better in Dynamic mode.
4x512MB double sided is better than 2x1024 double sided, as
the former config has twice as many memory banks as the latter
configuration.
It is not possible to write a closed form expression for performance.
How the memory can be driven, depends on its past history. For
example, if a memory page is already open, the next access to it
can be faster. If memory access is random, there can be a lot of
page opening and closing, to degrade performance. Thus, the
workload (memory access pattern and frequency of access) plays
a part in what the performance will be. Obviously, a memory
benchmark is an optimistic case, as the pattern is bursting and
open pages are fully utilized. Real programs don't do that.
On the issue of clock rate, overclocking is a slight winner over
tight CAS memory. A CAS3 DDR500 memory config with 1:1 CPU:Mem
ratio is faster than CAS2 DDR400 memory config with 5:4 CPU:Mem.
But, when running four sticks, the math becomes more cloudy,
because you don't know how far above DDR440 you would get with
4 double sided 512MB sticks.
On command rate, you can experiment. For example, some people
get away with changing 2T to 1T on the Athlon64 when using two
double sided sticks. I don't know anything about 875/865 and
how important the command rate setting is, or whether the user
can change it. I cannot find a register in the datasheet for
the 875 that controls command rate!
For PAT, it is really a crapshoot. Get a copy of CTIAW, or use
CTIAW as a search term on abxzone.com/forums/search.php . The
only condition that _might_ give you PAT, is a single stick of
CAS2 memory per channel, FSB800, DDR400, no overclock. (So the
Corsair TwinX mentioned below would fit the bill.)
Speaking in general terms, I like your plan of buying 2x1024.
These modules use 64Mx8 chips and are pretty expensive. Visit
this page and look at the specs for TWINX2048-3200C2
http://corsairmicro.com/corsair/xms.html
http://corsairmicro.com/corsair/products/specs/twinx2048-3200c2.pdf
If you want the stability of a stock system, buy your fast
processor, run it at FSB800, DDR400 (1:1 ratio), slap in
the 3200C2 1GB modules, run CTIAW, and there is the odd chance
you'll even get CTIAW to claim PAT is enabled.
If you are an overclocker, I cannot guarantee that someone won't
beat you with another memory config. If a module comes along
tomorrow, with lower capacitive loading, and DDR500+ performance,
four of those might beat you.
Here is some overclocker chatter to consider, posted very
recently.
http://www.abxzone.com/forums/showthread.php?t=80311
Intel offers this document, but with the results of the
article you site, their advice is now in doubt.
ftp://download.intel.com/design/chipsets/applnots/25273001.pdf
Whatever you decide, post back any significant results of
your testing
HTH,
Paul
